CoDeR is a new framework that constructs executable worlds using code and renders them visually with video generation models. It uses five roles: a concept interpreter, a creator that designs rules, executors that write code, artists that generate videos, and travelers that explore. This separation allows the world to maintain persistent state even when not observed, and supports long-term memory, open-ended interactions, autonomous evolution, and multi-agent scenarios. Experiments show state-of-the-art performance across multiple evaluation settings for these capabilities. The approach addresses key limitations of existing video-based world models that implicitly encode state and struggle with consistency and off-screen dynamics.
Executable code gives generative worlds persistent rules and evolving state
CoDeR separates world simulation from visual rendering, coordinating five roles to build executable dynamics and turn their evolving state into video.
Chinese Tech
Zixun Fang · Yawen Shao · Kai Zhu · Jie Xiao · Shihan Chen · Yu Liu · +4 more
USTC · TongYi Lab
Research Digest··2 min read
In one line
CoDeR combines explicit code-based world rules with diffusion video generation to enable open-ended world simulation.
Before you read this
Three ideas this paper builds onworld model for agentsA predictive environment model that agents use to simulate outcomes, enabling planning, reasoning, and long-horizon task execution.113 papers →code as actionA paradigm where actions are expressed as executable code, such as method calls or program synthesis, rather than direct manipulation.29 papers →multi-agent architectureA system where multiple AI agents collaborate, each handling specialized roles, communicating and coordinating to complete complex tasks.110 papers →
Read it as
Fang and colleagues present CoDeR, a world-modeling framework that represents rules, entities, and state explicitly in code rather than only through generated video.
The authors designed a system in which five complementary roles translate a high-level concept into structured rules, executable dynamics, and visual observations.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.